Method for determining vehicle flow on a road network

Abstract

The present invention relates to a method for determining the vehicle flow rate on a road segment of a road network. A model for determining the vehicle flow rate on a road segment is constructed as follows: for each road segment of a training road network, an ordered sequence is determined according to a direction of travel, comprising static descriptors of the training road segment and its neighbors; a short- and long-term memory neural network is trained on the ordered sequences and the result is concatenated with dynamic descriptors of the training road segments; a transform neural network is trained on these concatenations. Then, the vehicle flow rate on a road segment of a road network of interest is determined by applying the constructed model to static and dynamic descriptors of the road segment and to static descriptors of neighboring road segments. Figure 2A to be published

Description

Title of the invention: Method for determining vehicle flow on a road network technical field

[0001] The present invention relates to the field of determining vehicle traffic information for a road network. In particular, the invention relates to a method for determining vehicle flow on at least one section of road in a road network.

[0002] Today, metropolitan areas and road managers have travel modeling tools at their disposal to plan and simulate the impact of future traffic regulations and works, in order to reduce congestion or improve air quality. However, these models are very difficult to implement (requiring very expensive population survey data and a significant calibration effort), difficult to maintain (maintenance is delegated to consulting firms, making it difficult to develop the tool quickly), and difficult to update (population surveys are conducted every 5 to 10 years).

[0003] There is therefore a strong need for tools to determine vehicle flow rates within a transport network without having to use overly expensive and difficult-to-obtain data as input, such as mobility surveys, which are also static and not suitable for predicting rapid changes in mobility, or actual usage data, such as "floating car data" (FCD), which is dynamic but not necessarily available for all road segments of an urban network. Furthermore, these alternative tools should be easy to use for non-experts, quick to run in order to easily evaluate and compare several case studies, and finally, reliable by reproducing recent traffic count data on the road network under consideration as closely as possible. Previous technique

[0004] Most alternative approaches in the literature require one or more measurements on a given road segment to estimate the corresponding traffic flow. Their accuracy depends on the penetration rate of the measurements and their availability. Therefore, it would be impossible to estimate the flow on a road segment not covered by these measurements using these methods. Some documents propose extrapolations and spatio-temporal correlations to estimate vehicle flows on a road network even without complete measurement coverage. These require continuous updating of the measurements and are often time-intensive. Calculations and computer memory requirements, with matrices that grow exponentially with the road network. Their use on a large road network may therefore be limited.

[0005] Forecasting vehicle flow is a fundamental topic for the operation and real-time management of transportation systems. Applications range from road network monitoring and incident detection to the calculation of emissions and air quality. Numerous studies in the literature propose estimating traffic flow from counting loop measurements. They therefore exploit historical time series data to predict traffic flow over a certain time horizon (temporal extrapolations). The models proposed to address this problem are quite broad and range from autoregressive statistical models to machine learning models such as neural networks, support vector machines (SVMs), or random forests. Deep learning approaches have also been explored in the literature.These approaches often lead to a better understanding of temporal correlations compared to machine learning approaches. However, these methods are limited to temporal predictions without necessarily studying spatial extrapolation. Therefore, it is impossible to predict flow rates on road sections without flow measurements using these methods.

[0006] Other methods have been proposed to estimate traffic flow using other sensors. One method developed involves estimating road traffic flow from data obtained from surveillance videos. Other solutions involve statistical approaches and supervised learning approaches such as CNNs (Convolutional Neural Networks), GMMs (Gaussian Mixture Models), or Kalman filters to estimate the number and direction of vehicles. Other methods are based on estimating road traffic flow from noise measurements from acoustic beacons. This approach involves extracting features from the acoustic signal and correlating them with road traffic measurements using supervised learning methods.However, the method requires a continuous acoustic signal from a sensor located near a road segment to avoid background noise. Cellular data has also been used in the literature for traffic flow estimation. These approaches, however, have limited accuracy, which depends on the penetration rate and availability of the measurements. Therefore, it would be impossible to estimate vehicle flow on a road segment not covered by these measurements.

[0007] Other authors have proposed methods for estimating road traffic flow from various data and sensors present on a road network. Some proposed methods exploit FCD data and video surveillance data and / or counting loop measurements to estimate flow on a complete road network. The approaches can range from semi-supervised graph learning to geometric matrix completion and temporal spatial filters. This type of approach thus makes it possible to estimate flow on road segments even without measurements. However, this requires a more or less significant penetration of measurements on the road segments with continuous updating of the measurements. These approaches are also often computationally intensive, with matrices that grow exponentially with the size of the road network. Their application on a large road network may therefore be limited.

[0008] Data fusion techniques are also known, aimed at combining different sources of information, which can include sensor data, historical traffic data, meteorological data, etc. Finally, location data from smartphones (or smart phones, to use the French term for them) and navigation applications are also used to estimate traffic density and predict traffic conditions.

[0009] Most of the aforementioned methods require at least one measurement on a specific road segment to estimate the associated traffic flow. The accuracy of these methods depends closely on the penetration rate of available measurements and their availability. Consequently, these approaches prove incapable of estimating traffic flow on segments not covered by these measurements (this is referred to hereafter as spatial extrapolation).

[0010] However, document CN111292534A is known, which concerns a method for estimating traffic conditions on urban highways, using k-means classification and deep sequence learning to estimate traffic conditions across the entire highway network, even when real-time data are not available for certain road sections. The method involves dividing the highway network, modeling the data, preprocessing the data, performing cluster analysis, and applying a deep sequence learning model to provide traffic information. However, this method has the drawback of being limited to highway segments, which form a network of long, successive sequences, unlike urban segments, which have more significant and frequent intersections and bifurcations.

[0011] We also know of document CN 111222491A, which relates to a deep learning-based traffic evaluation system for estimating traffic flow using Characteristic image sequences and convolutional neural networks are used. The document aims to improve vehicle number detection based on the orthographic projection of their area in images. This method involves processing images obtained from cameras using deep learning. However, this method has the drawback of relying on the availability of traffic images (from cameras, for example) to deduce the flow. Furthermore, it is limited to locations where this information is available, unlike GIS (Geographic Information System) data, which is available virtually worldwide and easily accessible.

[0012] We also know of document JP2023082282A2, which concerns a travel time estimation method capable of predicting the travel time required for a vehicle using a highway to a predefined point using deep learning models. The proposed models first estimate the variations in upstream traffic volume in a continuous section of the highway, and then calculate the travel time based on these predictions, also using output information to improve the accuracy of the predictions. This method, however, appears to be specific to highways. Furthermore, the vehicle flow is calculated only upstream of the section of the highway being used, in order to estimate a travel time.

[0013] We also know of document IN202041039998A, which relates to a method for predicting road traffic flow, based on big data to generate large-scale traffic data and which applies machine learning, image processing, and deep learning algorithms to analyze this data, thus offering the possibility of predicting traffic flow and guiding autonomous vehicles for better real-time traffic management. The document requires real-time data from mobile application users, which is then analyzed and combined with other information from autonomous vehicle sensors to better guide them. However, this method has the drawback of requiring real-time data (from applications such as Google Maps™, for example), in addition to image and / or video data collected by autonomous vehicles.

[0014] We also know of document CN101593424A, which concerns a method for forecasting short-term traffic flow and which is based on three modules: a first historical average module, which divides the different dates of a year into three different types (working days, holidays, and weekends) and calculates the different traffic flows for each type based on historical flow data; a second neural network module that estimates traffic flows through data modeling; and finally, a third module that takes into account The method considers various traffic conditions and combines the output of the historical average module and the neural network module to predict short-term traffic flow. This allows for the prediction of future traffic flow on road segments based on available flow measurements on those same segments. In other words, this method enables temporal extrapolation but not spatial extrapolation.

[0015] EP22204301A1 is also known, which relates to a method for determining vehicle flow on a road segment, including for a road segment not equipped with a sensor. For this method, flow measurements are taken, using at least one fixed sensor, at at least one measurement point on a training road network. A vehicle flow model is then built, using machine learning, from macroscopic data of the training road network and the measurements. This model is then applied to the road segment under consideration. This method makes it possible to quickly estimate (without traffic simulation tools) the traffic flow on a large road network in the absence of prior measurements. However, this method does not guarantee the spatial continuity of the predicted flow between adjacent segments. Finally, this method is not sufficiently accurate, as will be shown in the second application example below.

[0016] The present invention overcomes these drawbacks. In particular, the present invention relates to a method for estimating vehicle flow on any road segment of a road network at any given time. The method according to the invention is based on the combination of two deep neural networks, one taking into account spatial information from a training road network, and the other taking into account temporal information on the same network. Furthermore, in the invention, the spatial information taken into account for deep learning concerns a road segment and the adjacent segments, which allows the neural network to learn spatial correlations between the characteristics of the segments and the associated traffic flow.In general, the method according to the invention makes it possible to increase the accuracy of vehicle flow estimates, to improve the spatial continuity of the predicted flow (in particular by taking into account adjacent road strands in strand "triplets"), to accelerate calculations and to reduce the memory footprint of the trained models. Summary of the invention

[0017] The present invention relates to a method for determining a vehicle flow rate on at least one road segment of a road network, by means of a training database comprising, for each training road segment, a plurality of training road segments of a training road network, at least a measure of vehicle flow, at least one static descriptor, at least one dynamic descriptor, and at least one static descriptor for learning road strands of said learning road network having a node in common with said learning road strand.

[0018] The method according to the invention comprises at least the following steps:

[0019] A) A model is constructed to determine a vehicle flow rate on a road segment of a road network based on static and dynamic descriptors of said road segment of said road network by applying at least the following sub-steps:

[0020] I) for each learning road strand of said plurality of learning road strands, an ordered sequence of static descriptors is determined relative to said learning road strand as a function of said static descriptors of said learning road strand and of said static descriptors or of a statistic of said static descriptors of said learning road strands having a node in common with said learning road strand, said ordered sequence of static descriptors being ordered according to a predefined direction of movement of said vehicles on said learning road strand;

[0021] II) a short-term and long-term memory neural network is trained on said plurality of ordered sequences of static descriptors relating to said plurality of learning road strands;

[0022] III) for each learning road strand of said plurality of learning road strands, a result of said short-term and long-term memory neural network for said learning road strand is concatenated with said dynamic descriptors of said learning road strand;

[0023] IV) a transformative neural network is trained on said plurality of concatenations of said result of said short-term and long-term memory neural network for said learning road strand with said dynamic descriptors of said learning road strand;

[0024] B) at least one static descriptor and at least one dynamic descriptor are obtained for said road strand of said road network, as well as at least one static descriptor for road strands having a node in common with said road strand;

[0025] C) by means of said at least one static descriptor and said at least one dynamic descriptor of said road strand of said road network, as well as of said at least one static descriptor for said road strands having a node in common with said road strand, and by means of said model to determine a vehicle flow on a road strand of a road network as a function of static and dynamic descriptors of said strand of said road network, said vehicle flow on said road strand of said road network is determined.

[0026] According to one embodiment of the invention, said learning base can be constructed by carrying out at least the following substeps for each learning road strand of said plurality of learning road strands of said learning road network:

[0027] i) At least one vehicle flow rate is measured on said training road segment by means of at least one fixed traffic measurement sensor;

[0028] ii) At least one static descriptor and at least one dynamic descriptor are obtained for said learning road strand, as well as a static descriptor for each of said road strands having a common node with said learning road strand.

[0029] According to one embodiment of the invention, said fixed traffic measurement sensor may be a counting loop.

[0030] According to one embodiment of the invention, said static descriptors of said at least one road strand of said road network, respectively of said learning road strands of said learning road network, may include at least one type, number of lanes, length and maximum speed of said road strand, respectively of said learning road strands.

[0031] According to one embodiment of the invention, said dynamic descriptors of said at least one road strand of said road network, respectively of said learning road strands of said learning road network, may consist of a sequence of average vehicle speeds for each hour of a day on said road strand, respectively on said learning road strands.

[0032] According to one embodiment of the invention, said statistic may be an average.

[0033] According to one embodiment of the invention, for said at least one road strand, for each of the said learning road segments of the said learning road network, we can determine the said ordered sequence of static descriptors relating to said road segment, respectively to said learning road segment, by forming a vector $ written: s=[ÿ, x, Z ]

[0034] where the vector x corresponds to said static descriptors of said road strand, respectively of said learning road strand, Y and Z are vectors of said statistics of said static descriptors of said road strands, respectively of said learning road strands, having a node in common with said road strand, respectively with said learning road strand, and located upstream and downstream of said road strand, respectively of said learning road strand, according to said predefined direction of traffic.

[0035] According to one embodiment of the invention, for said at least one road strand, respectively for each of said learning road strands, said result of said short-term and long-term memory neural network can be concatenated for said at minus one road strand, respectively for said learning road strand, with said dynamic descriptors of said at least one road strand, respectively for said learning road strand, in the following manner:

[0036] ST = [h, T]

[0037] where is a vector corresponding to the output of said short-term and long-term memory neural network for said at least one road strand, respectively for said at least one learning road strand, T is a vector of said dynamic descriptors of said at least one road strand, respectively of said learning road strand, in the form T= (rt: t GT), where Tt is one of said dynamic descriptors of said at least one road strand, respectively of said learning road strand, measured at time step r of a measurement period F.

[0038] According to one embodiment of the invention, infrastructure of said road network can be adapted according to said vehicle flow determined for said at least one road segment, and / or speed limits of said road network can be adapted according to said vehicle flow determined for said at least one road segment, and / or said vehicle flow determined for said at least one road segment can be displayed and / or an air quality estimated from said vehicle flow determined for said at least one road segment can be displayed.

[0039] In addition, the invention relates to a computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing the method as described above, when said program is executed on a computer.

[0040] Other features and advantages of the process according to the invention will become apparent from the following description of non-limiting examples of embodiments, with reference to the figures attached and described below. List of figures [Fig IA] [Fig IB] [Fig IC]

[0041] Figure 1A (respectively Figure 1B and Figure 1C) shows a curve representing the measured flow rate as a function of the time of day and a curve representing the flow rate as a function of the time of day determined by the method according to the invention for a first (respectively a second and a third) road segment of the validation database from the training road network. [Fig 2A]

[0042] Fig. 2A presents a map of the flow estimated by the method according to the invention on the main axes of the city of Lyon. [Fig 2B]

[0043] Fig. 2B presents an enlargement of part of the urban area of ​​the map in Fig. 2A. Description of the implementation methods

[0044] The present invention relates to a method for determining the flow of vehicles on at least one segment of a road network, preferably on the entire road network. In other words, the present invention aims to count the number of vehicles per unit of time that pass through a section of road.

[0045] According to one embodiment of the invention, the vehicle flow rate on the road segment considered may be, for example, an average flow rate or a maximum flow rate per unit of time, or more generally any statistic relating to a vehicle flow rate (minimum, standard deviation, median, etc.). The maximum vehicle flow rate reflects the capacity of a road segment. It is a physical parameter that determines the dimensions of the road segment.

[0046] According to one embodiment of the invention, the vehicle flow rate on the road segment in question can be a daily flow rate (for example, average or maximum), that is, the vehicle flow rate on the road segment in question over a period of one day. This can make it possible to predict the variation in flow rate from one day to the next, particularly due to the day of the week, the month, public holidays, school holidays, etc. Alternatively, the vehicle flow rate on the road segment in question can be an hourly flow rate (for example, average or maximum), that is, the vehicle flow rate on the road segment in question over a period of one hour. This can make it possible to predict the variation in flow rate within a day, particularly due to work schedules, school schedules, leisure schedules, etc.

[0047] Generally speaking, a "road network" is defined as a set of roads and paths within a predefined geographical area. This predefined geographical area can be a district of a city, a city, a group of municipalities, a department, etc. A segment of the road network is an elementary subdivision of the road network between two consecutive nodes of the road network. A road node is an endpoint of a road segment, which may correspond to an intersection, an obstacle, a change in road signs, etc. For example, a segment of the road network could be a road between two consecutive intersections, between two consecutive signs, between an intersection and a sign, or a section of highway between two consecutive exits, etc. Thus, we have a detailed breakdown of the road network and a model that is adapted to the road network without microscopic data.

[0048] In a classic but not limiting manner, the road network can be represented by a graph, called a road graph. The road graph is composed of a set of edges (also called arcs) and nodes, the nodes representing intersections, and the edges representing road segments between intersections. The graph A road graph can be obtained from an online mapping service (“web service”), for example Here™ (Here Apps LLC, Netherlands), which provides the graph edges as pure geometric objects. Preferably, the road graph is consistent with the road network (all physical connections between any two roads, and only those, are represented by the nodes of the graph), and as time-invariant as possible. Furthermore, the road graph can be simplified by excluding road segments such as dead ends, paths in parks, or cycle paths, depending on the type of vehicle considered (for example, for the embodiment of motor vehicles, cycle paths may be excluded).

[0049] The method according to the invention can be implemented for all types of vehicles: bicycles, motor vehicles, motorized two-wheelers, boats, hovercrafts, scooters, etc., as well as pedestrian travel. However, the method according to the invention is particularly suited to motorized vehicles, such as motor vehicles, trucks, buses, motorized two-wheelers, etc.

[0050] The method according to the invention requires measurements obtained by a fixed road traffic sensor. Each fixed road traffic sensor is positioned at a point (called a measurement point) on a road network (in this case, a training road network). Each fixed sensor allows for point (spatial) and temporal measurements of the number of vehicle passages (vehicle flow) and, optionally, the speed of the passing vehicles (traffic speed). The positions of the measurement points are identified within the road network. According to one embodiment of the invention, the fixed road traffic sensors can be cameras, radars, inductive loops, sensitive photoelectric cells detecting the interruption of a light beam, piezoelectric cables measuring the pressure exerted on the road surface, or any similar sensors.

[0051] According to a preferred embodiment of the invention, the fixed road traffic sensor corresponds to an inductive loop, also called a counting loop. Indeed, freely accessible databases of road flow measurements taken by counting loops exist. The implementation of the method according to the invention is thus simplified.

[0052] The method for determining vehicle throughput according to the invention can implement the following steps:

[0053] 1) Construction of the learning base

[0054] 1.1) Measurement of vehicle flow on a training road network

[0055] 1.2) Obtaining static and dynamic descriptors of a plurality of strands road workers of the road learning network

[0056] 2) Construction of a model to determine a vehicle flow rate on a strand of road network using the learning base

[0057] 2.1) Determination of a plurality of ordered static descriptor sequences

[0058] 2.2) Training an LSTM neural network on a plurality of sequences static descriptor orders

[0059] 2.3) Concatenation of the LSTM neural network result with the descriptors dynamics

[0060] 2.4) Training a neural network Transform on the concatenation of the Results of the LSTM neural network with dynamic descriptors

[0061] 3) Determination of vehicle flow on a section of a road network using the model for determining vehicle flow on a section of a constructed road network

[0062] 3.1) Obtaining static and dynamic descriptors of at least one road strand of the road network

[0063] 3.2) Determination of ordered sequences of static descriptors

[0064] 3.3) Application of the trained LSTM neural network

[0065] 3.4) Concatenation of the LSTM neural network result with the descriptors dynamics

[0066] 3.5) Application of the Transformer neural network trained on the concatenation of the Results of the LSTM neural network with dynamic descriptors

[0067] Steps 1) and 2) can be performed only once, and beforehand (in "offline" mode). In other words, once the model for determining vehicle flow on a segment of a road network has been built using the training dataset, this model can be applied one or more times. For example, if it is desired to determine the vehicle flow on several segments of a road network or the vehicle flow on the same road segment but at different times, only step 3 of the method according to the invention needs to be repeated.

[0068] Advantageously, the training set can be updated regularly (for example every month and at least every year), and generally as soon as new data are available, and then the model can be updated to determine a vehicle flow rate on a strand of a road network, by training this model on the enriched training set.

[0069] At least step 2) and / or step 3) can be implemented by computer means, in particular a computer, a processor or a calculator.

[0070] In the remainder of the application, the road network to which step 3) is applied is referred to as the "considered road network," the "prediction road network," or the "inference road network." It could also be called the "second road network." This is the road network for which the vehicle flow rate is to be determined. These notations distinguish it from the "training road network." or the "first road network," which is used for building the learning base. The "road network under consideration" may correspond to or include the "learning road network," or it may be a road network different from the "learning road network."

[0071] The steps of the process according to the invention are detailed below. 1) Building the learning base

[0072] During this step, the aim is to build a training dataset to train the model to determine a vehicle flow rate on at least one strand of a road network.

[0073] This step includes at least the two substeps described below.

[0074] 1.1) Measurement of vehicle flow on a training road network

[0075] During this sub-step, it is necessary to carry out measurements of the flow of vehicles on a training road network, using fixed road traffic sensors.

[0076] Preferably, vehicular flow measurements are taken at at least one point on a plurality of road segments of the training road network. Conventionally, the plurality of input data for a neural network increases the network's ability to represent the reality to be modeled, and this is even more true in deep learning. Preferably, the training set includes flow values ​​for at least 100 road segments of a training road network, and most preferably for at least 500 road segments of a training road network.

[0077] Preferably, vehicle flow measurements are repeated over time at a given point on each road segment of a plurality of road segments in the training road network, for example, every hour, preferably every minute, throughout the day. Such measurement frequency allows for the precise recording of traffic conditions over time, in order, for example, to determine peak hours and observe the various flow fluctuations during a day.

[0078] Advantageously, when measurements are repeated over time for a period exceeding one week, preferably one month, and preferably one year, a categorical descriptor (for example, a class representing the day of the year) can be associated with the measurement, allowing the measurements to be categorized according to a given temporal segmentation, for example, weekdays, weekends, holidays, etc. This can enable the neural network to learn to extrapolate temporally by incorporating a categorical descriptor specifying the chosen temporal segmentation. Thus, it will be possible to predict the throughput for any day of the year.

[0079] 1.2) Obtaining static and dynamic descriptors of a plurality of strands road workers of a training road network

[0080] In this substep, the aim is to obtain static and dynamic descriptors (or parameters, or attributes) for the plurality of road segments in the training road network, as well as static descriptors for the road segments that share a node with the training road segment. In other words, the training dataset is constructed by taking into account static descriptors for the training road segments, but also for neighboring, or adjacent, road segments. As will be described below, this helps to ensure spatial continuity of the predicted flow rate between adjacent segments.

[0081] A static descriptor relating to a strand of a road network (whether training or not) is understood to be a parameter relating to the strand of the road network that is invariant over time. The static descriptors of the strands of a road network according to the invention may also be called "macroscopic data of the road network" or "structural data" in the literature. Macroscopic data corresponds to information related to the road network, such as infrastructure, gradient, signage, etc. As will be described below, the invention takes into account the static descriptors relating to a road strand (whether training or not), as well as the static descriptors relating to its directly neighboring strands (having at least one node in common).

[0082] According to one aspect of the invention, the static descriptors of a strand of a road network (learning or not) can include the topography (i.e. the slope, length, turns, intersections, etc.) of the road strand, the type of road strand (e.g. urban road, motorway, etc.), the number of lanes of the road strand, the maximum speed of the road strand, the signage of the road strand, etc.

[0083] Preferably, the static descriptors of a strand of a road network (learning or not) include at least the type of road strand, the number of lanes of the road strand, the maximum speed of the road strand and the length of the road strand.

[0084] Preferably, the static descriptors of the segments of a road network (whether training or not) can be obtained using a geographic information system (GIS). Here™, Google Maps™, and OpenStreetMap™ are examples of geographic information systems. In general, GIS models a road network as a graph, where each segment is considered as a link between two nodes whose static descriptors are invariable. This representation is particularly well-suited to the implementation of the invention, since for each segment of a road network, its upstream and downstream segments are considered, as will be described below for the purposes of the invention.

[0085] By dynamic descriptor (or temporal descriptor or time-varying descriptor) relating to a strand of a road network (learning or not), This refers to a parameter related to a section of the road network that varies over time. It could be a time series showing the average speed of vehicles on the road section in question for each hour of a day. Such information can also be obtained via GIS.

[0086] As will be described later, the method according to the invention separately exploits the static and dynamic descriptors according to the invention, or in other words, the method according to the invention implements a learning architecture that separates the static descriptors from the dynamic descriptors. The separation of the static and dynamic descriptors makes it possible to use different models (in this case, recurrent neural networks) adapted to their specific characteristics. For example, as will be described below, so-called "attention" models are better suited to retaining memory of long temporal sequences, and are therefore better suited to dynamic descriptors. Similarly, as will be described below, LSTM-type neural networks make it possible to learn the spatial dependencies of relatively short sequences of static descriptors.

[0087] Thus, at the end of this step, the learning base according to the invention includes at least static and dynamic descriptors for a plurality of road strands of a learning road network, as well as static descriptors of the strands having a node in common with the road strands of the learning road network.

[0088] Advantageously, and for the purposes of step 2) of building the model, the training set can be divided into 2, preferably 3, subsets:

[0089] - a first subset, comprising for example between 70% and 85%, of 80% of the training set data is intended for training neural networks: this data will directly determine the estimation error measure of the networks and how to adjust the parameter values ​​of the neural network; this is the training set.

[0090] - a second subset, comprising for example between 15% and 30%, of 20% of the training dataset is reserved for validation during execution of the training process. This data is used to monitor the success of the training operation but does not directly influence the adjustment of network parameters; this is the validation dataset. If the results obtained are unsatisfactory, it may mean that the training dataset is not sufficiently representative and that it may be necessary to augment the training dataset with data covering a wider range of road traffic volumes by adding road segments.

[0091] - optionally but advantageously, a third subset, comprising a number of data points from the training set in a proportion equivalent to the validation database, intended to test learning performance at the end of learning; this is the test database.

[0092] 2) Construction of a model to determine a vehicle flow rate on a strand of a road network using the learning base

[0093] In this step, the aim is to construct a model to determine a flow of vehicles on a strand of a road network from static and dynamic descriptors of the strand of the road network, by training, on the learning basis constructed in step 1), a combination of two recurrent neural networks, composed of a short and long term memory neural network (known by the acronym LSTM, for "Long and Short Term Memory" in English) followed by a transformer neural network ("Transformer" in English).

[0094] The LSTM-type network belongs to the family of recursive neural networks (RNNs). A recursive neural network is an artificial neural network with recurrent connections. A recursive neural network consists of interconnected units (neurons) that interact non-linearly and for which there is at least one cycle in the structure. The units are linked by arcs (synapses) that have a weight. The output of a neuron is a non-linear combination of its inputs. Thus, RNNs allow for tracking a sequence of information. In particular, these networks are especially well-suited to sequential data because they use internal memory to retain information about previous time steps. However, recursive neural networks encounter the vanishing gradient problem when learning to memorize past events.

[0095] LSTM neural networks are improved variants of RNNs that address the vanishing gradient problem. Each computational unit is linked not only to a hidden state but also to a cell state that acts as memory. The transition occurs via constant-gain transfer, equal to 1. In this way, errors propagate to previous steps without the vanishing gradient phenomenon. The cell state can be modified through an input gate, which allows or blocks the update. Similarly, an output gate controls whether the cell state is communicated to the output of the LSTM unit. The most widespread version of LSTM also uses a gate that allows the cell state to be reset to zero ("forget gate").

[0096] The transformer network is also a variant of recurrent neural networks. Transformer networks are suitable for handling sequential data. However, unlike other RNNs, transformers do not require that the sequential data be processed in order. Thanks to this feature, the transformer allows for much greater parallelization. larger than RNNs, and therefore with reduced training times. Chained RNNs process tokens sequentially, maintaining a state vector that contains a representation of the data seen after each token. To process the nth token, the model combines the state representing the input sequence up to the (nl)th token with the information from the new token to create a new state representing the input sequence up to the nth token. Theoretically, the information from a token can propagate far back towards the beginning of the input sequence if, at each point, the state continues to encode information about the token. But in practice, this mechanism is imperfect: due to the gradient evanescence problem, the model state at the end of processing a long input sequence cannot recall precise information about the first tokens.To solve this problem, transformer-type networks introduce attention mechanisms. These mechanisms allow a model to directly look at, and extract, the state of any previous token in the sequence. The attention layer can access all previous states and weight them according to their relevance to the current token, thus providing salient information about distant tokens.

[0097] The nesting according to the invention of the two aforementioned recurrent neural networks is carried out according to steps 2.1 to 2.4 described below.

[0098] 2.1) Determination of a plurality of sequences of static descriptors ordered

[0099] During this substep, for each road strand of the plurality of road strands of the learning road network, an ordered sequence of static descriptors is determined relative to the learning road strand as a function of the static descriptors of the learning road strand and the static descriptors or a statistic of the static descriptors of the learning road strands having a node in common with the learning road strand, the ordered sequence of static descriptors being ordered according to a predefined direction of traffic of vehicles on the learning road strand.

[0100] In other words, according to the invention, the ordered sequence of static descriptors comprises the static descriptors or a statistical descriptor of the static descriptors of the learning road strands having a node in common with the road strand considered and located upstream of the strand considered according to the predefined direction of traffic, followed by the static descriptors of the road strand considered, themselves followed by the static descriptors or the statistical descriptors of the road strands having a node in common with the road strand considered and located downstream of the strand considered according to the predefined direction of traffic.

[0101] Such an ordered sequence makes it possible to take into account the static descriptors of the road strands adjacent or even neighboring to the road strand considered, moreover according to a direction of traffic.

[0102] According to a preferred embodiment of the invention, the statistics of the static descriptors of road strands having a node in common with the road strand under consideration can be an average. This makes it possible to take into account average trends of the static descriptors on adjacent road strands.

[0103] The sequence of static descriptors ordered according to the invention associated with a road strand can be represented in the form of a vector. More precisely, consider a road strand having static descriptors represented by a vector xg P^- Let Ye pÆ xd be matrices relating to the k and k strands (k and k' can be greater than or equal to 1) located upstream and downstream (relative to the direction of flow) of strand x, each upstream and downstream strand having d descriptors respectively.

[0104] According to an embodiment in which the static descriptors of the strands adjacent to the strand under consideration are taken into account in the ordered sequence of static descriptors according to the invention in the form of a statistic (for example, a mean), the vector x can be concatenated with a vector Y and a vector Z comprising respectively the statistic (by type of descriptor) of the static descriptors of the matrices Y and Z. This yields an ordered sequence (represented as a vector), hereafter called a triplet, which can be written as:

[0105] s = z]-

[0106] Thus, in this design, the vector $ represents an ordered sequence of static descriptors, hereafter denoted, relating to a road strand, such that S = (sp: pe 3 xd )•

[0107] According to an embodiment in which each of the static descriptors of the strands adjacent to the strand in question is taken into account in the ordered sequence of static descriptors according to the invention, the ordered sequence according to the invention can be written:

[0108] S = [nf ...,Y1K, .... YDL .... YDK, x,Zlf ....ZUT, ZDl ... ; , ZDK'] with D varying from 1 to d, K varying from 1 to k, and K' varying from 1 to k', and $ ep(*+*'+i)xd. Thus, in this conception, the vector 5' represents an ordered sequence of static descriptors, hereafter denoted sp, relating to a road strand, such that S = Gp ; p & (k + k + 1) xd )•

[0109] At the end of this sub-step, we obtain an ordered sequence of static descriptors for each road strand of the plurality of road strands of the learning road network according to the invention.

[0110] 2.2) Training an LSTM neural network on a plurality of sequences static descriptor orders

[0111] During this substep, an LSTM neural network is trained on the plurality of ordered sequences of static descriptors determined in the previous substep.

[0112] Subsequently, the following notation is used for the activation vectors of the linked LSTM neural network: - at the front door: ip; - at the exit gate: °p; - at the door of oblivion: fp; - to the state of the cell: cp; - at the exit door: hp.

[0113] According to one embodiment of the invention, the activation vectors of the LSTM neural network, for each static descriptor sp, can be written as follows:

[0114] ip= <7( Wi - [hp.^ fy)

[0115] Op= &(WO - [hp.1, + bo)

[0116] fp = a(Wr [hp.b + bf)

[0117] cp= fp-cp^ + ip• tanh(wc• [h^, + b^

[0118] _ 0p.tanh(cp)

[0119] such that p is the spatial step (or the descriptor index) in the triplet -S, y, and VE. c are weight vectors relative to the input gate ip, the output gate °p, the forget gate fp, and the cell state cp, respectively, and b^bo, bf, and bc are bias vectors relative to the input gate ip, the output gate °p, the forget gate fp, and the cell state cp, respectively, and where 17 is a sigmoid function defined by . The weight vectors and bias vectors are learned during the training process.

[0120] The above equations are valid only for a single spatial step. Thus, if the triplet is a sequence of N elements, these equations are calculated N times.

[0121] 2.3) Concatenation of the result of the LSTM neural network with the dynamic descriptors

[0122] During this substep, for each road strand of the plurality of road networks in the training base, the result (or output) of the LSTM determined in step 2.2) for the road strand in question is concatenated with the dynamic descriptors of that road strand.

[0123] In other words, during this step, the output h of the LSTM determined for the ordered sequence of static descriptors S of a strand x is concatenated with the dynamic descriptors of the strand x, denoted T - (rt: te P), according to the following formula:

[0124] ST = [h, T]

[0125] such that T is the measurement period of the dynamic descriptors considered (for example a duration of 24 hours).

[0126] 2.4) Training a neural network: Transforming on concatenation of the LSTM neural network output with dynamic descriptors

[0127] During this substep, the transformer neural network is trained on the concatenation of the result of the LSTM neural network with the dynamic descriptors of each road strand, referred to hereafter as the "Transformer input". In conventional terms, this means that the input, denoted ST, of the Transformer is fed into the encoder part of the transformer.

[0128] According to one embodiment of the invention, the encoder can be composed of a stack of identical encoding layers with the same series of operations. Each encoder can mainly contain a multi-head attention layer, a positional encoding layer, and a linear layer connected with a ReLU (Rerectified Linear Unit) type activation function.

[0129] According to one implementation, in the attention layer, the transformer can learn three weight matrices: the Wq query weight matrix, the key matrix called the attention head. These weight matrices contain parameters that are optimized during training. These parameters are used to weight the input values. The attention, denoted A, for the entire sequence can be expressed in matrix representation for the whole ST sequence as follows:

[0130] n (QKT\ K, V) = softmax -7^ . V \ /

[0131] where qe pwx^ and yep* x each row of the keys of the The queries Q and values ​​U are a vector representation of elements in the sequence, and softmax is a function to convert a vector of real numbers into a probability distribution. The softmax function normalizes the attention scores, ensuring that the attention given to each component of the input sequence to the transformer is proportional to its relative importance compared to the other components in the sequence.

[0132] According to one embodiment of the invention, during the transformer drive, the ST sequence composed of static and dynamic descriptors k and the weight matrix of Wy values, 1 set (yy jy ST = / \ is transmitted all at once in the first Vf ^2' ......^3xd+r / The encoder block then processes the input, and the information is then passed on to its successor, that is, the next encoder. This process is repeated until all N encoding blocks have processed the transformer's input. An encoded representation of ST is then obtained. In this context, recurrent neural network (RNN) models are limited in their ability to use information from elements observed far in the past within the sequence. More generally, they have more difficulty linking sequential information that is distant from one another. Attention, on the other hand, allows for the efficient and rapid linking of each element in one sequence to all elements in another sequence, as well as to all other elements within the same sequence.

[0133] Thus, at the end of this step, we obtain a model to determine a flow of vehicles on a strand of a road network, from static and dynamic descriptors of the strand of the road network, the model comprising a combination of two recurrent neural networks, specifically an LSTM and a Transformer, trained on a learning basis.

[0134] 3) Determination of vehicle flow on at least one segment of a road network using the model to determine a vehicle flow rate on a strand of a constructed road network

[0135] In this step, the model built on a training road network as described in step 2) is applied to at least one road segment of a second road network, in order to determine the vehicle flow rate on that road segment.

[0136] More specifically, during this step, the following sub-steps are carried out:

[0137] 3.1) Obtaining static and dynamic descriptors of at least one strand road network

[0138] During this sub-step, the aim is to obtain static and dynamic descriptors relating to said at least one road strand of the road network under consideration, as well as static descriptors relating to the road strands of the second road network having a common node with said at least one road strand under consideration and being located upstream and / or downstream (with respect to a predefined direction of traffic) of said at least one road strand under consideration.

[0139] It is quite clear that this sub-step corresponds in every respect to sub-step 1.2) described above, except that it is applied here to at least one road strand of the road network under consideration and to its adjacent strands, and no longer to a plurality of road strands of the learning road network and to their adjacent strands.

[0140] 3.2) Determination of ordered sequences of static descriptors

[0141] During this substep, for at least one road segment of the road network under consideration, an ordered sequence of static descriptors is determined, said ordered sequence of static descriptors being ordered according to a direction of traffic of said vehicles on said road segment, said ordered sequence of static descriptors comprising an average of the static descriptors of the road segments having a node in common with the road segment under consideration and located upstream of the segment under consideration according to said predefined direction of traffic, followed by the static descriptors of the road segment under consideration, followed by a statistic (for example an average) of the static descriptors of the road segments having a node in common with the road segment under consideration and located downstream of the segment under consideration according to said predefined direction of traffic.

[0142] It is quite clear that this sub-step corresponds in every respect to sub-step 2.1) described above, except that it is applied to said at least one road segment of the road network under consideration, and no longer to a plurality of road segments of the learning road network.

[0143] 3.3) Application of the trained LSTM neural network

[0144] During this substep, for at least one road segment of the road network under consideration, the LSTM neural network trained in step 2.2) is applied to at least the ordered sequence of static descriptors of the road segment determined in step 3.2).

[0145] 3.4) Concatenation of the result of the LSTM neural network with the dynamic descriptors

[0146] During this substep, for said at least one road segment of the road network under consideration, the result (or output) of the LSTM determined in step 3.3) for the road segment under consideration is concatenated with the dynamic descriptors of this road segment.

[0147] It is quite clear that this sub-step corresponds in every respect to sub-step 2.3) described above, except that it is applied to said at least one road segment of the road network under consideration, and no longer to the plurality of road segments of the learning road network.

[0148] 3.5) Application of the Transformer neural network trained on the concatenation of the LSTM neural network output with the dynamic descriptors

[0149] During this substep, the transformative neural network is applied to the concatenation of the result of the LSTM neural network with the dynamic descriptors of said at least one road strand.

[0150] Thus, the method according to the invention is based on taking into account road segments and their immediate surroundings, moreover according to a direction of travel (upstream and downstream segments). This description allows for a threefold increase in input data and the incorporation of the direction of travel into the structure of the descriptors. Furthermore, This allows us to account for the fact that road segments are intrinsically linked. Indeed, the traffic state of a road segment depends on the traffic state of neighboring segments located upstream and downstream. Furthermore, this allows us to consider that two segments with the same static descriptors will not necessarily exhibit the same flow variations.

[0151] Furthermore, using the LSTM as a model for extracting the spatial features of a road strand obtained via static descriptors allows the triplet sequence to be considered as a static series, thus preserving the continuity of flow between adjacent strands. To predict the time series representing the flow, the LSTM output (containing a dense representation of the spatial features respecting the strand order) is concatenated with the dynamic descriptors. This creates a sequence twice as long as the LSTM output. This sequence then constitutes the input to the transformer, which allows for the capture of long-term patterns. Its output can be followed by linear layers and ReLU activation functions to predict the flow.Thus, the deep learning method implemented in the process according to the invention allows, via an LSTM, the capture of local features, and via a Transformer, the capture of long-term patterns. Furthermore, the present invention makes it possible to obtain flow rate estimates for any territory without prior measurements being taken on that territory.

[0152] According to one embodiment of the invention, based on the vehicle flow rate on at least one segment of a given road network as described above, one can: - adapt the infrastructure of the road network according to said vehicle flow determined for said at least one road segment, for example by creating new lanes, modifying signage, etc. From the flow, public authorities and public works companies can indeed determine the roads with a high vehicle flow and adapt the roads to the users; - and / or adapt the speed limits of the road network according to said vehicle flow determined for said at least one road section, for example via illuminated signs according to the estimated flow downstream of said signs, in order to improve traffic flow; - and / or display said vehicle flow determined for said at least one road segment or a parameter determined from said vehicle flow such as air quality, for example via a display on the dashboard of vehicles using the road network, on a portable, standalone device, such as a geolocation device (such as a GPS), a mobile phone (such as a smartphone), on a website, etc. The display may take the form of a note, a color code, or a thickness of representation of the road. A road network user may, depending on the predicted flow for the road segment in question, choose to take a route not including that road segment; - and / or display an estimated air quality based on the flow rate determined as described above, for example via a connected device (a sign) in a street or group of streets (neighborhood). Information on the vehicular flow rate of a road segment can indeed allow for an estimation of the air quality near that road segment.

[0153] Thus, the present invention can be viewed as a virtual vehicle flow sensor, in that it allows for the simulation of a vehicle flow measurement. A virtually unlimited number of these virtual sensors can therefore be deployed at low cost to obtain an overview of the flow rate over a given area.

[0154] Furthermore, the invention relates to a computer program product downloadable from a communication network and / or stored on a computer-readable medium (e.g., an embedded computer) and / or executable by a processor. This program includes program code instructions for implementing the method as described above, in particular steps 2) and / or 3) described above, when the program is executed on a computer. Examples

[0155] The characteristics and advantages of the method according to the invention will become clearer upon reading the application example below.

[0156] A training dataset of road strands for which road flow rates are known was used. For the construction of the model according to the invention, 70% of the strands are used to train the model according to the invention and the remaining 30% are used for validation. First example#

[0157] The first example consists of a comparison between the actual hourly flow rate (measured) and the flow rate estimated by the method according to the invention for three road segments from the validation database. Figure 1A (respectively Figure 1B and Figure IC) shows a MES curve representing the flow rate D (in number of vehicles per hour, denoted veh / h) measured as a function of the time H of the day (in hours) and an INV curve representing the flow rate D (in number of vehicles per hour, denoted veh / h) as a function of the time H of the day (in hours) determined by the method according to the invention for a first (respectively a second and a third) road segment from the validation database derived from the training road network. It can be seen in these figures that the model according to the invention is capable of predicting a significant range of flow rate values ​​(values ​​between 300 and 900 veh / h).

[0158] Furthermore, the INV curves determined by the method according to the invention follow the trends of the measured MES curves, thus making it possible to predict congestion times corresponding to the morning and evening rush hours (around 8 a.m. and 6 p.m.). Second example# The second example consists of a comparison of the throughput determined by the prior art method described in the aforementioned document EP22204301A1 and by the method according to the invention. Table 1 presents a comparison of the absolute error, the mean of the GEH statistic, the median value of the GEH statistic, and the percentage of GEH greater than 5, determined on the one hand by the method according to the invention (column INV) and on the other hand by the prior art method (AA). The GEH statistic (from the initials of its creator, Geoffrey E. Havers) is a formula used in traffic engineering (particularly traffic forecasting and modeling) to compare two traffic volumes (typically, a series of measured actual traffic and a series of traffic generated by modeling). It can be written as:

[0159]

[0160] ^“GEH y j+y Where are the actual measurements and y are the values ​​estimated by a process. It is classically considered that if _ ^veh / hj and mo'ns ^es flow rates predicted have a GEH greater than $ then the modeled flow values are considered accurate. It can be observed that the method according to the invention has made it possible to considerably reduce the values ​​of these quantities, thus demonstrating the superior extrapolation capabilities of deep learning for predicting throughput. In particular, the absolute error value has been reduced by a factor greater than 3, the average GEH is less than 5 and therefore meets the conditions, and finally, the percentage of GEH greater than 5 has been halved. [Tables 1] INV AA Absolute error (yeh / h) 59.67 203.4 Mean GEH 4.48 8.8 Median GEH 3.06 7.1 % of GEH > 5 30.52 63.1 Third example In the third example, the method according to the invention is applied to the city of Lyon (France) for a weekday, during peak hours (6 p.m.). The input data

[0161] The model's static and dynamic descriptors are derived from GIS services such as Here Maps. Figure 2A shows a map of the flow rate D (in number of vehicles per hour, represented in grayscale, the unit of which is denoted veh / h) estimated by the method according to the invention on the main roads of the city of Lyon, and Figure 2B shows an enlargement of a portion located in the urban area of ​​the map in Figure 2A. It can be observed in these figures that the peripheral segments (i.e., the segments located on the outskirts of the city) exhibit a significant flow rate compared to the urban segments (i.e., the segments located in the city center). This corresponds well to the observed reality of road traffic. Furthermore, the spatial continuity of the flow rate is preserved. Indeed, the order of magnitude of the flow rate remains the same for consecutive segments where there are no intersections or crossroads.Furthermore, the flow rate prediction by the method according to the invention is obtained in approximately l / 4h on an Intel Core i7 @2.60 GHz processor with 16 GB of memory, i.e. a duration divided by two compared to the method according to the aforementioned prior art.

Claims

1. Demands A method for determining vehicle flow on at least one road segment of a road network, characterized in that at least the following steps are carried out using a training dataset comprising, for each training road segment, a plurality of training road segments of a training road network, at least one vehicle flow measurement, at least one static descriptor, at least one dynamic descriptor, and at least one static descriptor for training road segments of said training road network having a node in common with said training road segment: A) A model is constructed to determine a vehicle flow rate on a road segment of a road network from static and dynamic descriptors of said road segment of said road network by applying at least the following sub-steps: I) for each learning road strand of said plurality of learning road strands, an ordered sequence of static descriptors is determined relative to said learning road strand as a function of said static descriptors of said learning road strand and of said static descriptors or of a statistic of said static descriptors of said learning road strands having a node in common with said learning road strand, said ordered sequence of static descriptors being ordered according to a predefined direction of movement of said vehicles on said learning road strand; II) a short-term and long-term memory neural network is trained on said plurality of ordered sequences of static descriptors relating to said plurality of learning road strands; III) for each learning road strand of said plurality of learning road strands, a result of said short-term and long-term memory neural network for said learning road strand is concatenated with said dynamic descriptors of said learning road strand; IV) A transforming neural network is trained on the said plurality of concatenations of the said result of the said short- and long-term memory neural network for the said road strand a) learning with said dynamic descriptors of said learning road segment; B) at least one static descriptor and at least one dynamic descriptor of said road segment of said road network are obtained, as well as at least one static descriptor for road segments having a node in common with said road segment; C) by means of said at least one static descriptor and said at least one dynamic descriptor of said road segment of said road network, as well as said at least one static descriptor for said road segments having a node in common with said road segment, and by means of said model for determining a vehicle flow on a road segment of a road network as a function of static and dynamic descriptors of said segment of said road network, said vehicle flow on said road segment of said road network is determined.

2. A method according to claim 1, wherein said learning base is constructed by carrying out at least the following substeps for each learning road strand of said plurality of learning road strands of said learning road network: i. At least one vehicle flow rate is measured on said learning road strand by means of at least one fixed traffic measurement sensor; ii. At least one static descriptor and at least one dynamic descriptor are obtained for said learning road strand, as well as a static descriptor for each of said road strands having a common node with said learning road strand.

3. Method according to claim 2, wherein said fixed traffic measurement sensor is a counting loop.

4. A method according to any one of the preceding claims, wherein said static descriptors of said at least one road strand of said road network, or of said learning road strands of said learning road network, comprise at least one type, number of lanes, length and maximum speed of said road strand, or of said learning road strands.

5. A method according to any one of the preceding claims, wherein said dynamic descriptors of said at least one road strand of said road network, respectively of said learning road strands said road learning network, consist of a sequence of average vehicle speeds for each hour of a day on said road segment, respectively on said road learning segments.

6. A method according to any one of the preceding claims, wherein said statistic is an average.

7. A method according to any one of the preceding claims, wherein, for said at least one road strand, respectively for each of said learning road strands of said learning road network, said ordered sequence of static descriptors relating to said road strand, respectively to said learning road strand, is determined by forming a vector $ written: S = XZ] Where the vector x corresponds to said static descriptors of said road strand, respectively to said learning road strand, Y and Z are vectors of said statistics of said static descriptors of said road strands, respectively to said learning road strands, having a node in common with said road strand, respectively with said learning road strand, and located upstream and downstream of said road strand, respectively to said learning road strand, according to said predefined direction of traffic.

8. A method according to any one of the preceding claims, wherein, for said at least one road strand, or for each of said learning road strands, the output of said short- and long-term memory neural network for said at least one road strand, or for said learning road strand, is concatenated with said dynamic descriptors of said at least one road strand, or for said learning road strand, in the following manner: ST = [A, T] where h is a vector corresponding to the output of said short- and long-term memory neural network for said at least one road strand, or for said at least one learning road strand, and T is a vector of said dynamic descriptors of said at least one road strand, or for said learning road strand, in the form T = (vt : te F) where Tt is one of said dynamic descriptors of said at least one road strand, or for said learning road strand. learning route, measured at time step * of a measurement period T.

9. Product computer program downloadable from a communication network and / or stored on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing the method according to any one of the preceding claims, when said program is executed on a computer.

10. A method according to any one of claims 1 to 8, wherein road network infrastructure is adapted according to said vehicle flow determined for said at least one road segment, and / or speed limits are adapted according to said vehicle flow determined for said at least one road segment, and / or said vehicle flow determined for said at least one road segment is displayed and / or an air quality estimated from said vehicle flow determined for said at least one road segment is displayed.

Images